Deep Unsupervised Domain Adaptation Research Articles

Label-Free Poisoning Attack Against Deep Unsupervised Domain Adaptation

Deep unsupervised domain adaptation (UDA) has significantly boosted the performance of deep models on different domains by transferring knowledge from a source domain to a target domain. However, its robustness against adversarial attacks has not been explored due to the challenges of highly non-convex deep models and different data distribution. In this paper, we give the first attempt to analyze the vulnerability of deep UDA and propose a label-free poisoning attack (LFPA), which injects poisoning data into the training data to mislead adaptation between the two domains without ground truth in target domain. Specifically, we design an unsupervised adversarial loss as the attack goal, in which the pseudo-labels are used to approximate the ground-truth. Since retraining the model will gradually degrade the attack performance, we also add a regularization term to the unsupervised loss, which eliminates negative interactions between the training goal and the attack goal. To accelerate the craft of poisons, we select influential samples as the initial poisons and propose a fast reverse-mode optimization method which updates poisons according to the approximate truncated gradients. Experimental results on multiple state-of-the-art deep UDA methods demonstrate the effectiveness of the proposed LFPA and the high sensitivity of UDA to poisoning attacks.

Enhanced Sample Self-Revised Network for Cross-Dataset Facial Expression Recognition

Recently, cross-dataset facial expression recognition (FER) has obtained wide attention from researchers. Thanks to the emergence of large-scale facial expression datasets, cross-dataset FER has made great progress. Nevertheless, facial images in large-scale datasets with low quality, subjective annotation, severe occlusion, and rare subject identity can lead to the existence of outlier samples in facial expression datasets. These outlier samples are usually far from the clustering center of the dataset in the feature space, thus resulting in considerable differences in feature distribution, which severely restricts the performance of most cross-dataset facial expression recognition methods. To eliminate the influence of outlier samples on cross-dataset FER, we propose the enhanced sample self-revised network (ESSRN) with a novel outlier-handling mechanism, whose aim is first to seek these outlier samples and then suppress them in dealing with cross-dataset FER. To evaluate the proposed ESSRN, we conduct extensive cross-dataset experiments across RAF-DB, JAFFE, CK+, and FER2013 datasets. Experimental results demonstrate that the proposed outlier-handling mechanism can reduce the negative impact of outlier samples on cross-dataset FER effectively and our ESSRN outperforms classic deep unsupervised domain adaptation (UDA) methods and the recent state-of-the-art cross-dataset FER results.

Deep Unsupervised Domain Adaptation with Time Series Sensor Data: A Survey

Sensors are devices that output signals for sensing physical phenomena and are widely used in all aspects of our social production activities. The continuous recording of physical parameters allows effective analysis of the operational status of the monitored system and prediction of unknown risks. Thanks to the development of deep learning, the ability to analyze temporal signals collected by sensors has been greatly improved. However, models trained in the source domain do not perform well in the target domain due to the presence of domain gaps. In recent years, many researchers have used deep unsupervised domain adaptation techniques to address the domain gap between signals collected by sensors in different scenarios, i.e., using labeled data in the source domain and unlabeled data in the target domain to improve the performance of models in the target domain. This survey first summarizes the background of recent research on unsupervised domain adaptation with time series sensor data, the types of sensors used, the domain gap between the source and target domains, and commonly used datasets. Then, the paper classifies and compares different unsupervised domain adaptation methods according to the way of adaptation and summarizes different adaptation settings based on the number of source and target domains. Finally, this survey discusses the challenges of the current research and provides an outlook on future work. This survey systematically reviews and summarizes recent research on unsupervised domain adaptation for time series sensor data to provide the reader with a systematic understanding of the field.

Joint bi-adversarial learning for unsupervised domain adaptation

An important challenge of unsupervised domain adaptation (UDA) is how to sufficiently utilize the structure and information of the data distribution, so as to exploit the source domain knowledge for a more accurate classification of the unlabeled target domain. Currently, much research work has been devoted to UDA. However, existing works have mostly considered only distribution alignment or learning domain invariant features by adversarial techniques, ignoring feature processing and intra-domain category information. To this end, we design a new cross-domain discrepancy metric, namely joint distribution for maximum mean discrepancy (JD-MMD), and propose a deep unsupervised domain adaptation learning method, namely joint bi-adversarial learning for unsupervised domain adaptation (JBL-UDA). Specifically, JD-MMD measures cross-domain divergence in terms of both discrepancy and relevance by preserving cross-domain joint distribution discrepancy, as well as their class discriminability. Then, with such divergence measure, JBL-UDA models with two learning modalities, one is founded by the bi-adversarial learning from domains and classes implicitly, while the other explicitly addresses domains and classes alignment via the JD-MMD metric. Besides, JBL-UDA explores structural prior knowledge from data classes and domains to generate class-discriminative and domain-invariant representations. Finally, extensive evaluations exhibit state-of-the-art accuracy of the proposed methodology.

A Review of Single-Source Deep Unsupervised Visual Domain Adaptation.

Large-scale labeled training datasets have enabled deep neural networks to excel across a wide range of benchmark vision tasks. However, in many applications, it is prohibitively expensive and time-consuming to obtain large quantities of labeled data. To cope with limited labeled training data, many have attempted to directly apply models trained on a large-scale labeled source domain to another sparsely labeled or unlabeled target domain. Unfortunately, direct transfer across domains often performs poorly due to the presence of domain shift or dataset bias. Domain adaptation (DA) is a machine learning paradigm that aims to learn a model from a source domain that can perform well on a different (but related) target domain. In this article, we review the latest single-source deep unsupervised DA methods focused on visual tasks and discuss new perspectives for future research. We begin with the definitions of different DA strategies and the descriptions of existing benchmark datasets. We then summarize and compare different categories of single-source unsupervised DA methods, including discrepancy-based methods, adversarial discriminative methods, adversarial generative methods, and self-supervision-based methods. Finally, we discuss future research directions with challenges and possible solutions.

Feature Based Bearing Fault Detection With Phase Current Sensor Signals Under Different Operating Conditions

To avoid sudden motor failures, which might lead to unexpected downtimes of entire production plants, predictive machinery maintenance is of great importance. Since bearingsare the only connection between rotor and stator and for thisreason, among others, more than 40 percent of all motor failures are reasoned by bearing defects, bearing defect detectionis of particular importance for predictive maintenance of motor failures. In the past, mainly external vibration sensorswere used for fault detection. However, in recent years, therehas been increasing research into using phase currents forbearing damage detection, which has the advantage of savingadditional sensors and thus costs. The practical use of phasecurrents is justified by the fact that bearing damage is accompanied by eccentricities between the rotor and stator, whichaffect the magnetic flux and thus also the phase currents. Forthese reasons, we focus on bearing damage detection usingphase currents.We show in the paper that variations of the rotational shaftfrequency, torque loads and radial forces applied to the bearings have a strong influence of the applicability of a defectdetection model trained with phase current data. This means,a defect detection model, which was trained with labeled datacollected under a fixed operating condition is not able to makereliable classifications if the operating conditions differ fromthe training conditions. This is due to discrepancies in feature distributions between the different operating conditions.However, it is of utmost importance for practical applicability that a fault detection model is able to make reliable classifications under a variety of previously unknown operating conditions.To solve this problem, we present a deep learning approachfor diminishing the influence of different operating conditionsby utilizing unsupervised domain adaptation. As input datafor the neural network, we propose to extract hand-craftedfeatures, both from time- and frequency domain, from thephase current signals. This greatly reduces training time andresource utilization compared to training a neural network asfeature extractor on the raw time series sensor signals. Thebasic idea is then to train a neural network with labeled datafrom the source domain, i.e. data of one defined operatingcondition, and at the same time consider unlabeled data ofthe new, previously unknown, target operating condition inorder to achieve higher classification accuracies for this operating condition as well. Thus, labeled data is only necessaryfrom the source domain to train the Softmax classificationlayer of the deep neural network. As the features computedfrom the differing working conditions are the same, domaindiscrepancy can be reduced in an unsupervised manner byweight sharing and formulating a specific loss term (we useMulti-Layer Maximum Mean Discrepancy) that captures thediscrepancy between the different operating conditions. Byapplying this additional loss term to the backpropagation process, discrepancies between the feature means of source- andtarget domain are reduced iteratively.The effectiveness of the proposed method is evaluated with abenchmark dataset from the University of Paderborn, which iswell known in the bearing damage community. The proposedfeature based deep unsupervised domain adaptation methodsignificantly increases classification accuracies for all transfer scenarios. With the best of our knowledge, it is the firsttime one deals with phase current based domain adaptationon statistical fault-features.

Deep visual unsupervised domain adaptation for classification tasks: a survey

Learning methods are challenged when there is not enough labelled data. It gets worse when the existing learning data have different distributions in different domains. To deal with such situations, deep unsupervised domain adaptation techniques have newly been widely used. This study surveys such domain adaptation methods that have been used for classification tasks in computer vision. The survey includes the very recent papers on this topic that have not been included in the previous surveys and introduces a taxonomy by grouping methods published on unsupervised domain adaptation into five groups of discrepancy-, adversarial-, reconstruction-, representation-, and attention-based methods.

A study of the effects of negative transfer on deep unsupervised domain adaptation methods

Intelligent systems driven by deep learning have become relevant in real-world applications with the increasing availability of technology and data. However, real-world settings require effective and robust deep learning models that are able to deal with unforeseen samples and a variety of data distributions. Recently, Unsupervised Domain Adaptation (UDA) for deep learning models (D-UDA) addresses such limitations by transferring knowledge from a labeled source domain to an unlabeled target domain, reducing the dataset shift between domain distributions. However, despite recent advances in D-UDA, current works have not been focused on studying specific cases in the distribution shifts under which D-UDA methods can ensure that transfer is helpful, avoiding a ‘negative transfer’ risk. In this paper, we present a study about the effect of different cases of negative transfer over the most popular and recent D-UDA methods reported in the literature. For this, we evaluate the accuracy performance of D-UDA methods over different scenarios containing different types of distribution shifts. Experimental results show that specific cases of distribution shifts generate negative transfer over the evaluated D-UDA methods. From this study, we provide some insights to select and design robust D-UDA methods in intelligent systems.

Transfer channel pruning for compressing deep domain adaptation models

Deep unsupervised domain adaptation has recently received increasing attention from researchers. However, existing methods are computationally intensive due to the computational cost of CNN (Convolutional Neural Networks) adopted by most work. There is no effective network compression method for such problem. In this paper, we propose a unified Transfer Channel Pruning (TCP) approach for accelerating deep unsupervised domain adaptation (UDA) models. TCP is capable of compressing the deep UDA model by pruning less important channels while simultaneously learning transferable features by reducing the cross-domain distribution divergence. Therefore, it reduces the impact of negative transfer and maintains competitive performance on the target task. To the best of our knowledge, TCP is the first approach that aims at accelerating deep unsupervised domain adaptation models. TCP is validated on two benchmark datasets – Office-31 and ImageCLEF-DA with two common backbone networks – VGG16 and ResNet50. Experimental results demonstrate that TCP achieves comparable or better classification accuracy than other comparison methods while significantly reducing the computational cost. To be more specific, in VGG16, we get even higher accuracy after pruning 26% floating point operations (FLOPs); in ResNet50, we also get higher accuracy on half of the tasks after pruning 12% FLOPs.

Deep multi-Wasserstein unsupervised domain adaptation

In unsupervised domain adaptation (DA), 1 aims at learning from labeled source data and fully unlabeled target examples a model with a low error on the target domain. In this setting, standard generalization bounds prompt us to minimize the sum of three terms: (a) the source true risk, (b) the divergence between the source and target domains, and (c) the combined error of the ideal joint hypothesis over the two domains. Many DA methods – especially those using deep neural networks – have focused on the first two terms by using different divergence measures to align the source and target distributions on a shared latent feature space, while ignoring the third term, assuming it is negligible to perform the adaptation. However, it has been shown that purely aligning the two distributions while minimizing the source error may lead to so-called negative transfer. In this paper, we address this issue with a new deep unsupervised DA method – called MCDA – minimizing the first two terms while controlling the third one. MCDA benefits from highly-confident target samples (using softmax predictions) to minimize class-wise Wasserstein distances and efficiently approximate the ideal joint hypothesis. Empirical results show that our approach outperforms state of the art methods.